Learning in a HUGE topic in artificial intelligence with major sub-disciplines for areas such as pattern recognition, reinforcement learning, learning by example, learning from analogies, and so on. Besides the type of learning, there is the question of when and where learning occurs (is the robot learning my preferences? learning to walk? or it is learning everything all the time such as lifelong learning satirized in The Complete Roderick. In robotics, topics are more constrained, and the purpose of learning falls into three categories, each with a different category of learning algorithms. One is learning about objects or events or how to hear and speak. This category of learning is referred to as pattern recognition and “deep learning” artificial neural networks have been a major breakthrough in classification. A major problem is learning to recognize when something is different- that that “orange” is really a “new type of orange”, a tangerine. That is not pattern recognition, which says “which of the 5 things I know about is this most likely to be” but rather the new term problem. Most algorithms are told to learn N things and can’t notice that there are really N+1 types of objects.  A second category is learning sequences of actions, like skills or strategies (we often call strategies “policies”, and this is usually done with reinforcement learning. The third area is learning intent, such as learning by demonstration to learn what the human wants the robot to do or learning what the person wants in a more social setting (like an assistant robot)​

Index:

• Overview: Science Robotics: Robot Learning in Science Fiction; Science Robotics: Computer Vision and Machine Learning in Science Fiction; A Brief History of Computer Vision and Machine Learning

• Deep Learning: Science Robotics: Computer Vision and Machine Learning in Science Fiction, A Brief History of Computer Vision and Machine Learning

• Learning by analogy: The Secret Life of bots

• Lifelong Learning: The Scarecrow’s Boy, The Complete Roderick, Short Circuit, Enthrian

• New Term Problem: The Scarecrow’s Boy

• Learning from demonstration (counter-example): Q.U.R.

• Reinforcement learning: Callahan and the Wheelies

For further reading:

• Chapter 16: Learning, Introduction to AI Robotics, 2nd Edition, R. Murphy,  MIT Press 2019 (available Aug)

• Springer Handbook of Robotics, 2nd Edition, Springer, 2016

• https://www.technologyreview.com/s/612002/a-plan-to-advance-ai-by-exploring-the-minds-of-children/ on Dr. Josh Tenenbaum’s work on lifelong learning by duplicating how children learn over time